Apprentice

Apprentice is an AI that learns physical work by watching. It tracks what happens in a video with SAM 3.1, uses Muse Spark to turn the tracks into measured physics and a step-by-step procedure, runs that procedure on a simulated robot that verifies its own work with the same eyes, and in coach mode flags the exact step a person got wrong. Two principles run through it: every number is checked against physics or a second observation, and every model call is priced against a hard budget. It measures gravity from a video to 0.11% median error across 50 clips; it watches one teacher video and performs the task in 60 brand-new scenes at 100% (vs 67% for copying the demo exactly); and as a coach it catches 21/21 injected mistakes with zero false alarms on clean attempts. Headline finding: asked to measure gravity from the same clips, a frontier model answers 9.81 every time (48% median error) - it recalls the constant; Apprentice measures it. The whole project cost $9.19.